Medical consumable illegal use identification method based on big data

Abstract

The invention relates to the technical field of medical consumables, in particular to a method for identifying illegal use of medical consumables based on big data, comprising the following steps: step 1, counting use of consumables; step 2, data acquisition; step 3, data cleaning; step 4, performing feature engineering; step 5, macroscopic analysis; step 6, performing mesoscopic analysis; step 7, microanalysis; according to the method, by constructing a risk early warning model based on big data and machine learning, illegal behaviors such as virtual high price, overuse and virtual remembering and excessive remembering occurring in the actual use process of the orthopedic medical consumables are recognized, the illegal behaviors of overuse of medical security funds in the illegal behaviors are mined, and the risk early warning effect of the medical consumables is improved. Through macroscopic analysis based on a time sequence algorithm, mesoscopic analysis based on a clustering algorithm and microscopic analysis based on an abnormal point detection algorithm, orthopedic medical consumable use data are comprehensively screened, and violation modes in the orthopedic medical consumable use data are locked.

Description

Technical Field

[0001] The present invention relates to the technical field of medical consumables, and in particular to a method for identifying illegal use of medical consumables based on big data. Background Art

[0002] At present, the identification of illegal use of orthopedic medical consumables is mainly completed through the following steps:

[0003] (1) Data screening rules: The information team, in cooperation with the policy team and the medical team, formulates screening rules, extracts suspicious data, and analyzes logical relationships based on reported complaint clues, common violation behavior characteristics, and local medical regulations. It also provides verification direction and data support for other inspections based on the screening results.

[0004] (2) Review and investigate medical records: Combined with the relevant content of the "Data Screening Rules", review key medical records, and expand the scope of medical records and conduct extended inspections in a timely manner after discovering problems. Focus on reviewing the patient's chief complaint, medical history records, auxiliary examinations, surgical records, etc., and combine with relevant disease diagnosis and treatment standards and guidelines to comprehensively determine whether there is excessive diagnosis and treatment, abuse of consumables, split hospitalization, etc.; review surgical records, consumables usage lists (registration forms), consumables disinfection records, consumables barcodes, and fee lists, and focus on comparing whether the information such as the number of consumables used, specifications, models, and manufacturers is consistent, and comprehensively determine whether there is consumables substitution, false records, split charges, etc.

[0005] (3) Review of imaging records: The medical team reviews the patient's preoperative and postoperative imaging records and other materials to verify the number and location of implanted consumables such as screws and plates. They also compare the information with the surgical records, consumable usage list (registration form), consumable barcodes, and fee list to verify whether the relevant information is consistent (focusing on the following situations: the number of consumables charged is greater than the number of implants, the charged consumables are not implanted consumables, and consumables such as Kirschner wires are used but not charged). Comprehensively determine whether there is any substitution of consumables, false records, or split charges.

[0006] (4) Check the purchase, sales and inventory data. The financial team, in combination with the relevant content of the "Data Screening Rules", retrieves the consumables inventory list, accompanying documents, purchase invoices, financial statements, accounting vouchers, warehouse entry and exit records, disinfection records, sales data, payment vouchers and other materials, and counts the actual inventory of consumables on site. It also classifies and checks the consumables purchase, sales and inventory data according to the types, specifications, manufacturers, prices, etc. of consumables, and then compares them with the corresponding consumables charging data to check whether the relevant data is consistent, whether the relevant bills are authentic, whether the purchase, sales and inventory ratio meets the local medical requirements, whether the financial accounts meet the management requirements, and whether the "zero markup" and "centralized procurement" policies are strictly implemented. And whether there are contradictions such as "inconsistency between accounts and actuals" and "multiple uses of one code", so as to comprehensively judge whether there are consumables substitution, false records, split charges, repeated charges, etc.

[0007] (5) Question key personnel. The medical team will conduct targeted questioning and investigation of relevant personnel in accordance with the law based on the "medical record review" and "data comparison". First, focus on questioning billing personnel, medical staff, consumables management personnel, etc. in turn to fully understand the work processes and responsibilities of relevant personnel. Second, focus on questioning medical staff to fully understand the operation process and methods of relevant surgeries. Third, for violations, relevant staff can be invited to work together.

[0008] In summary, existing methods and approaches are highly dependent on manual labor. Faced with increasingly numerous regulatory tasks, it is difficult to achieve full coverage, precision and efficiency in supervision. It is urgent to use big data + artificial intelligence technology to realize intelligent, full coverage and penetrating supervision. Summary of the invention

[0009] The purpose of the present invention is to provide a method for identifying illegal use of medical consumables based on big data to solve the problems raised in the above-mentioned background technology.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] The method for identifying illegal use of medical consumables based on big data includes the following steps:

[0012] Step 1: Count the usage of consumables: Count the related indicators of consumables usage and make Pareto analysis to determine the key monitoring list;

[0013] Step 2, data collection: extracting medical institution, department, physician, consumables, inventory, DRG / DIP information from data sources such as diagnosis and treatment database, regulatory department database, medical institution procurement database, and patient medical record database;

[0014] Step 3, data cleaning: processing missing values ​​and outliers, completing data format conversion, deduplication, type verification, and standardization;

[0015] Step 4, feature engineering: select and transform massive data so that its numerical regularity can be easily captured by the algorithm and reduce the impact of noise data on analysis accuracy;

[0016] Step 5, macro analysis: use a variety of time series analysis methods to do long-term trend analysis;

[0017] Step 6, meso-level analysis: Use a variety of extreme analysis, cluster analysis and other methods to find out the directionality and focus on the key points;

[0018] Step 7, micro-analysis: conduct abnormal detection and analysis on each type of consumables on the key monitoring list to find out the suspicious points;

[0019] Step 8, model adjustment: adjust steps 1 to 7 according to the actual model operation results to improve the model performance in an orderly manner.

[0020] Preferably, in the step of counting the use of consumables, the settlement amount of all medical consumables in the past three years is counted, and the Pareto analysis method is used to find out the types of consumables that meet the characteristics of large amount involved, large usage, and high unit price. The statistical caliber of the types refers to the national three-level classification catalog of medical consumables;

[0021] After analysis, it was found that in the past three years, the settlement amount of the top 30 medical consumables in the city accounted for 38% of the total settlement amount of all consumables, and formed a Pareto inflection point, thus being included in the key monitoring list.

[0022] Preferably, in the data acquisition step, the following data is acquired:

[0023] (1) Settlement lists and detailed items in the medical transaction database are used as the main data;

[0024] (2) Medical institution information, physician information, natural person information, and disease group information in the medical transaction database as dimensional data;

[0025] (3) Historical law enforcement and penalty information in the regulatory database is used as one of the label data;

[0026] (4) The consumables procurement and inventory information in the medical institution procurement database serves as the main data.

[0027] Preferably, in the data cleaning step, noise is eliminated by the following method:

[0028] (1) Remove missing values: For missing values, review the dictionary value of the field and process it by filling, deleting, or interpolating;

[0029] (2) Duplicate value processing: For each field’s duplicate values, review their meaning and delete necessary records;

[0030] (3) Data standardization / normalization: For continuous fields, standardization / normalization is performed to make them more consistent with model input requirements;

[0031] (4) Type conversion: For each field, convert it to a numeric or enumeration type to meet the model input requirements;

[0032] (5) Character processing: For each character field, delete special symbols, spaces, line breaks and invalid characters that affect data quality;

[0033] (6) Consistency check: Perform consistency check on the fields in each form that are used to connect or express the same meaning to ensure that their field names and field dictionary values ​​are consistent.

[0034] Preferably, in the feature engineering step, data features are extracted or enhanced mainly by the following methods:

[0035] (1) Data discretization: For some continuous fields, which are mainly expressed in segmented values ​​in business, discretization processing is performed;

[0036] (2) Sample balance: For data with labels, examine the ratio of each type of label and use the SMOTE algorithm to increase or decrease the amount of data corresponding to certain types of labels so that the amount of data corresponding to each type of label is close.

[0037] (3) Data encoding: Encode all categorical fields, such as natural encoding, One-Hot encoding, etc.

[0038] (4) Field segmentation: Split the merged fields. For example, split the “year, month, and day” field into “year”, “month”, and “day”.

[0039] Preferably, in the macro analysis step, the analysis is performed by a time series analysis method, and the analysis object is a variety of consumables usage indicators, mainly the usage quantity and cost, and the time statistical caliber is as high as the month and as low as the hour, so as to form a time series of indicators ranging from dozens to thousands for each type of consumables;

[0040] According to the characteristics of each time series, select appropriate analysis methods, including but not limited to the following methods:

[0041] (1) Autoregressive Integrated Moving Average (ARIMA) model: For time series analysis of indicators with "month" as the statistical caliber, the cycle, fluctuation, steady state and other information are extracted to determine the three parameters p, d, and q. After fitting the data, the values ​​of multiple points in the future can be predicted, and the confidence interval is set to no less than 90%;

[0042] (2) Prophet model: For time series analysis of indicators with "day" as the statistical caliber, the trend, discontinuity points, and fluctuation cycle information are extracted to determine important parameters of the model such as growth, checkpoints, and seasonal_mode. After fitting the data, the values ​​of multiple points in the future can be predicted, and the confidence interval is set to be no less than 90%;

[0043] (3) LSTM model based on attention mechanism: For the time series analysis of indicators with "hours" as the statistical caliber, the nonlinear correlation between the points in the specified time window is extracted and expressed in the form of normalized weights;

[0044] Through the multi-head attention mechanism, the changing patterns of consumables indicators at different scales can be accurately captured;

[0045] Through training, the specific structure of the model, the parameters of each layer, and the parameters of each neuron are determined, and the values ​​of multiple points in the future are predicted, with the confidence interval limited to no less than 90%.

[0046] Preferably, in the observation analysis step, the analysis is performed by extreme analysis and clustering algorithm, including:

[0047] (1) Extreme analysis: Use big data technology to conduct statistical analysis on the number of consumables used, screen cases where the number of one-time uses exceeds the threshold X, and focus on them. If the number of such cases is large, then proceed to the next step of cluster analysis; if the number is small, then turn to the correlation analysis of other items prescribed at the same time;

[0048] (2) Cluster analysis: K-means, DBSCAN, and BIRCH clustering algorithms were used to analyze the directivity, focusing on the three dimensions of medical institutions, doctors, and consumables manufacturers. The analysis was mainly based on statistical indicators such as type, quantity, number of people, and number of visits;

[0049] When payment is made according to DRG / DIP, additional directional analysis of disease groups is performed to obtain directional analysis in four dimensions: disease groups, medical institutions, doctors, and consumables manufacturers, and key marking is performed.

[0050] Preferably, in the micro-analysis step, the unreasonable usage range of the consumption and cost of each type of consumables is determined by an abnormal point analysis method with the consumable type as the core; the algorithms used are mainly isolation forest, single-class support vector machine, and local anomaly factor method;

[0051] (1) Isolation Forest: An efficient unsupervised learning algorithm for outlier detection. The basic principle is to recursively and randomly select data features and split points to split the data set into multiple subsets until only one data point remains in each subset. Since the number of abnormal data is small and the difference from normal data is large, fewer steps are required when they are isolated, so they can be identified faster. The steps are as follows:

[0052] Step 1, random initialization: randomly select the number of trees in the isolation forest;

[0053] Step 2, construct a single isolated number: for each isolated number, randomly select a data point from the data set as the root node of the tree; randomly select a feature and a split value on this feature; divide the data point into two subsets according to the split value, and each subset will serve as the root node of the left and right subtrees;

[0054] Step 3, for each non-leaf node: repeat Step 2, randomly select features and split values ​​until the stopping condition is met, which is either the preset maximum tree depth or the number of samples in the node is less than a certain threshold;

[0055] Step 4, repeat Step 1 to Step 3 to complete the construction of multiple trees;

[0056] Step 5, for each sample point, calculate its depth on each tree, and then calculate the average depth as the outlier score of the sample; outliers usually have shorter paths and therefore smaller depths;

[0057] Step 6, training: Use training data for training, and judge whether the model training is reasonable and appropriate by statistical anomaly scores, visualization, the structure of isolated trees, and the decision path of isolated trees;

[0058] Step 7, prediction: assign a prediction label to the new data point, -1 indicates abnormality and 1 indicates normality.

[0059] (2) Single-class support vector machine: an unsupervised learning algorithm widely used in the field of anomaly detection; its core idea is to maximize the distance from normal data points to the decision boundary, similar to the maximization of the margin in the support vector machine (SVM), so as to find the normal patterns in the data and mark the data points that are significantly different from these patterns as anomalies. The steps are as follows:

[0060] Step 1, set the kernel function to be used, usually selected from linear, polynomial, sigmoid and RBF kernels;

[0061] Step 2, set the parameters n and g, where n is a value between 0 and 1, indicating the allowable misclassification ratio, that is, the expected ratio of normal samples that are misclassified as abnormal, and g defines the influence range of a single training sample, usually used with the Gaussian kernel function;

[0062] Step 3, training: use the prepared data and set parameters to train the single-class support vector machine model;

[0063] Step 4, Evaluation: The model calculates a score for each sample, which indicates the distance from the sample to the decision boundary. Usually, samples with scores below a certain threshold are considered abnormal. The decision boundary is observed by visualization to adjust the model parameters to appropriate values.

[0064] Step 5, determine a threshold to distinguish normal and abnormal samples: this threshold can make a certain proportion of known normal samples be classified as normal scores as the threshold;

[0065] (3) Local anomaly factor: A density-based unsupervised anomaly detection method. The core idea is to identify anomalies by comparing the local density deviation of a data point with its neighbors. If the local density of a data point is significantly lower than that of its neighbors, then the point may be an anomaly. The density ratio of the data point to its k nearest neighbors is calculated to determine the anomaly score. This ratio is called the local anomaly factor. The steps are as follows:

[0066] Step 1, select the number of neighbors parameter k: Determine a parameter k, which represents the k nearest neighbors of each data point; this parameter is one of the most important hyperparameters in the algorithm and needs to be adjusted according to the specific data set;

[0067] Step 2, calculate K distance: For each data point P, calculate its distance to the kth neighbor among its k nearest neighbors. This distance is called the K distance of point P.

[0068] Step 3, determine the K neighborhood: use the K distance to determine the K neighborhood of each data point, that is, the set of all points whose distance from the data point P is not greater than K distance;

[0069] Step 4, calculate the reachable distance RD: For each neighbor O of the data point P, calculate the reachable distance from O to P, defined as max(K-distance(O), dist(P,O)), where dist(P,O) is the actual distance between P and O;

[0070] Step 5, calculate the local reachability density LRD: For each data point P, calculate the inverse of the average of the reachability distances of all its neighbors as the local reachability density of the point; the local reachability density reflects the local density of point P;

[0071] Step 6, calculate the local anomaly factor LOF: For each data point P, calculate its LOF score, which is the average of the ratios of the local reachability density of its k nearest neighbors to the local reachability density of P. This score reflects the degree of anomaly of point P relative to its neighbors;

[0072] Step 7: Sort all data points according to the LOF score, then set a threshold based on business requirements or data characteristics, and identify data points with LOF scores higher than the threshold as outliers.

[0073] Preferably, in the model adjustment step, the following is achieved by adjusting parameters of different algorithms and integrating algorithms of the same type:

[0074] (1) Parameter adjustment: When using algorithms to identify violations, there is often a lack of historical data with labels. For the three anomaly detection algorithms, Isolation Forest, One-Class Support Vector Machine, and Local Anomaly Factor, they have strong white-box properties, and the intermediate calculation process and results are easy to visualize. Their distinct statistical characteristics, decision boundaries, and decision paths can be observed. Cross-checking can be achieved through manual observation and statistical testing, so as to correct the results, reverse the model to adjust the parameters, and improve performance.

[0075] For ARIMA, Prophet, and LSTM models, as time series analysis algorithms, various point-by-point error indicators are used to evaluate model performance, and the training-evaluation process is embedded in the TPE optimization algorithm to achieve automatic optimization;

[0076] (2) Algorithm integration: The results of training and prediction using a single algorithm are often incomplete or inaccurate. The three anomaly detection algorithms, Isolation Forest, Single-Class Support Vector Machine, and Local Anomaly Factor, have completely different principles, but the results are all to mark abnormal data, so they are integrated; the results are fused through voting and weighting methods.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] The present invention constructs a risk warning model based on big data and machine learning to identify violations such as inflated prices, excessive use, false accounting and over-invoicing that occur in the actual use of orthopedic medical consumables, and discovers violations such as excessive use of medical insurance funds. Through macro analysis based on time series algorithms, meso analysis based on clustering algorithms, and micro analysis based on outlier detection algorithms, the use data of orthopedic medical consumables is comprehensively screened to identify violation patterns. The model can be used for diagnostic analysis in scenarios such as medical compliance inspections, commercial medical insurance claims, internal management of medical institutions, and administrative inspections by law enforcement agencies. DETAILED DESCRIPTION

[0079] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0080] The present invention provides a technical solution:

[0081] A method for identifying illegal use of medical consumables based on big data, characterized in that it comprises the following steps:

[0082] Step 1: Count the usage of consumables: Count the related indicators of consumables usage and make Pareto analysis to determine the key monitoring list;

[0083] Step 2, data collection: extracting medical institution, department, physician, consumables, inventory, DRG / DIP information from data sources such as diagnosis and treatment database, regulatory department database, medical institution procurement database, and patient medical record database;

[0084] Step 3, data cleaning: processing missing values ​​and outliers, completing data format conversion, deduplication, type verification, and standardization;

[0085] Step 4, feature engineering: select and transform massive data so that its numerical regularity can be easily captured by the algorithm and reduce the impact of noise data on analysis accuracy;

[0086] Step 5, macro analysis: use a variety of time series analysis methods to do long-term trend analysis;

[0087] Step 6, meso-level analysis: Use a variety of extreme analysis, cluster analysis and other methods to find out the directionality and focus on the key points;

[0088] Step 7, micro-analysis: conduct abnormal detection and analysis on each type of consumables on the key monitoring list to find out the suspicious points;

[0089] Step 8, model adjustment: adjust steps 1 to 7 according to the actual model operation results to improve the model performance in an orderly manner.

[0090] In the step of counting the use of consumables, this scheme mainly counts the settlement amount of all medical consumables in the past three years, and uses the Pareto analysis method to find out the types of consumables that meet the characteristics of large amounts involved, large usage, and high unit prices. The statistical caliber of the types refers to the national three-level classification catalog of medical consumables. After analysis, in the past three years, the settlement amount of the top 30 medical consumables in the city accounted for 38% of the total settlement amount of consumables, and formed a Pareto inflection point, so they were included in the key monitoring list.

[0091] In the data acquisition step of the scheme described above, the following data are mainly acquired:

[0092] (1) The settlement list (full fields) and detailed items (full fields) in the medical transaction database are used as the main data;

[0093] (2) Medical institution information (full field), physician information (full field), natural person information (full field), and disease group information (full field) in the medical transaction database are used as dimensional data;

[0094] (3) Historical law enforcement and penalty information (all fields) in the supervision database as one of the label data;

[0095] (4) The consumables procurement (all fields) and inventory information (all fields) in the medical institution procurement database serve as the main data.

[0096] In the data cleaning step, the scheme described here mainly eliminates noise by the following methods:

[0097] (1) Remove missing values: For missing (blank) values, review the dictionary value of the field and process it by filling (median, mode, etc.), deleting, or interpolating;

[0098] (2) Duplicate value processing: For each field’s duplicate values, review their meaning and delete necessary records;

[0099] (3) Data standardization / normalization: For continuous fields, standardization / normalization is performed to make them more consistent with model input requirements;

[0100] (4) Type conversion: For each field, convert it to a basic type such as numeric type or enumeration type to meet the model input requirements;

[0101] (5) Character processing: For each character field, delete special symbols, spaces, line breaks and other invalid characters that affect data quality;

[0102] (6) Consistency check: Perform consistency check on fields in each form that are used to connect or express the same meaning to ensure that their field names, field dictionary values, etc. are consistent.

[0103] In the feature engineering step, the scheme described here mainly uses the following methods to extract or enhance data features:

[0104] (1) Data discretization: For some continuous fields, which are mainly expressed in segmented values ​​in business, discretization processing is performed;

[0105] (2) Sample balance: For data with labels, examine the ratio of each type of label and use algorithms such as SMOTE to increase or decrease the amount of data corresponding to certain types of labels so that the amount of data corresponding to each type of label is close.

[0106] (3) Data encoding: Encode all categorical fields, such as natural encoding, One-Hot encoding, etc.

[0107] (4) Field segmentation: Split the merged fields. For example, split the “year, month, and day” field into “year,” “month,” and “day.”

[0108] In the macro analysis step, the scheme described in this paper mainly uses time series analysis methods for analysis. The analysis objects are various consumables usage indicators, mainly usage quantity and cost. The time statistical caliber is as high as the month and as low as the hour, so as to form a time series of indicators ranging from dozens to thousands for each type of consumables. According to the characteristics of each time series, select the appropriate analysis method. Including but not limited to the following methods:

[0109] (1) Autoregressive Integrated Moving Average (ARIMA) model: For time series analysis of indicators with "month" as the statistical caliber, the cycle, fluctuation, steady state and other information are extracted to determine the three parameters p, d, and q. After fitting the data, the values ​​of multiple points in the future can be predicted, and the confidence interval is set to no less than 90%;

[0110] (2) Prophet model: For time series analysis of indicators with "day" as the statistical caliber, extract information such as trends, discontinuities, and fluctuation cycles to determine important parameters of the model such as growth, checkpoints, and seasonal_mode. After fitting the data, the values ​​of multiple points in the future can be predicted, and the confidence interval is set to be no less than 90%;

[0111] (3) LSTM model based on attention mechanism: For the time series analysis of indicators with "hours" as the statistical caliber, the nonlinear correlation between the points in the specified time window is extracted and expressed in the form of normalized weights. Through the multi-head attention mechanism, the change pattern of indicators of consumables at different scales can be accurately captured. Through training to determine the specific structure of the model, parameters of each layer, parameters of each neuron, etc., the values ​​of multiple points in the future can be predicted, and the confidence interval is limited to no less than 90%.

[0112] In the meso-analysis step, the scheme described here mainly uses extreme analysis and clustering algorithms for analysis. It mainly includes:

[0113] (1) Extreme analysis: Use big data technology to conduct statistical analysis on the number of consumables used, screen cases where the number of one-time uses exceeds the threshold X, and focus on them. If the number of such cases is large, then proceed to the next step of cluster analysis; if the number is small, then turn to the correlation analysis of other items prescribed at the same time;

[0114] (2) Cluster analysis: Use clustering algorithms such as K-means, DBSCAN, and BIRCH to analyze the directionality, focusing on the three dimensions of medical institutions, doctors, and consumables manufacturers. The analysis is mainly based on statistical indicators such as type, quantity, number of people, and person-times. When payment is made according to DRG / DIP, add the directionality analysis of the patient group to obtain the directionality of the four dimensions of patient group, medical institutions, doctors, and consumables manufacturers, and mark them as key points.

[0115] In the micro-analysis step, the scheme mainly determines the unreasonable usage range of various consumables and costs through an abnormal point analysis method with consumable types as the core. The algorithms used are mainly isolation forest, single-class support vector machine, local anomaly factor and other methods.

[0116] (1) Isolation Forest: An efficient unsupervised learning algorithm mainly used for outlier detection. Its basic principle is to recursively and randomly select data features and split points to split the data set into multiple subsets until only one data point remains in each subset. Since the number of abnormal data is small and the difference is large from normal data, fewer steps are required when being isolated, so it can be identified faster. The steps are as follows:

[0117] Step 1: Random initialization. Randomly select the number of trees in the isolation forest;

[0118] Step 2: Construct a single isolated tree. For each isolated tree, randomly select a data point from the data set as the root node of the tree. Randomly select a feature and a split value on this feature. Divide the data point into two subsets according to the split value, and each subset will serve as the root node of the left and right subtrees;

[0119] Step 3: For each non-leaf node, repeat Step 2 and randomly select features and split values ​​until the stopping condition is met (the preset maximum tree depth is reached or the number of samples in the node is less than a certain threshold);

[0120] Step 4: Repeat Step 1 to Step 3 to complete the construction of multiple trees;

[0121] Step 5: For each sample point, calculate its depth on each tree, and then calculate the average depth as the outlier score of the sample. Outliers usually have shorter paths and therefore smaller depths;

[0122] Step 6: Training. Use the training data for training and judge whether the model training is reasonable and appropriate by statistically analyzing the anomaly score, visualization, the structure of the isolated tree, and the decision path of the isolated tree.

[0123] Step 7: Prediction. Assign a prediction label to the new data point, -1 indicates abnormality and 1 indicates normality.

[0124] (2) Single-class support vector machine: an unsupervised learning algorithm widely used in the field of anomaly detection. Its core idea is to maximize the distance from normal data points to the decision boundary, similar to the maximization of the margin in the support vector machine (SVM), so as to find the normal patterns in the data and mark the data points that are significantly different from these patterns as anomalies. The steps are as follows:

[0125] Step 1: Set the kernel function to be used, usually selected from linear, polynomial, sigmoid and RBF kernels;

[0126] Step 2: Set the parameters n and g, where n is a value between 0 and 1, indicating the allowable misclassification ratio, that is, the expected ratio of normal samples that are misclassified as abnormal, and g defines the influence range of a single training sample, usually used with the Gaussian kernel function;

[0127] Step 3: Training. Use the prepared data and set parameters to train a single-class support vector machine model;

[0128] Step 4: Evaluation. The model calculates a score for each sample, which indicates the distance from the sample to the decision boundary. Usually, samples with scores below a certain threshold are considered abnormal; the decision boundary is observed through visualization to adjust the model parameters to appropriate values;

[0129] Step 5: Determine a threshold to distinguish normal and abnormal samples. This threshold can make a certain proportion of known normal samples be classified as normal scores as the threshold.

[0130] (3) Local anomaly factor: A density-based unsupervised anomaly detection method. Its core idea is to identify anomalies by comparing the local density deviation between a data point and its neighbors. If the local density of a data point is significantly lower than that of its neighbors, then the point may be an anomaly. The density ratio of the data point and its k nearest neighbors is calculated to determine the anomaly score. This ratio is called the local anomaly factor. The steps are as follows:

[0131] Step 1: Select the number of neighbors parameter k. Determine a parameter k, which represents the k nearest neighbors of each data point. This parameter is one of the most important hyperparameters in the algorithm and needs to be adjusted according to the specific data set;

[0132] Step 2: Calculate the K distance. For each data point P, calculate the distance from it to the kth neighbor among its k nearest neighbors. This distance is called the K distance of point P.

[0133] Step 3: Determine the K neighborhood. Use the K distance to determine the K neighborhood of each data point, that is, the set of all points whose distance from the data point P is not greater than K distance;

[0134] Step 4: Calculate the reachable distance RD. For each neighbor O of data point P, calculate the reachable distance from O to P, defined as max(K-distance(O), dist(P,O)), where dist(P,O) is the actual distance between P and O;

[0135] Step 5: Calculate the local reachability density LRD. For each data point P, calculate the inverse of the average of the reachability distances of all its neighbors as the local reachability density of the point. The local reachability density reflects the local density of point P;

[0136] Step 6: Calculate the local outlier factor LOF. For each data point P, calculate its LOF score, which is the average of the ratios of the local reachability density of its k nearest neighbors to the local reachability density of P. This score reflects the degree of outlier of point P relative to its neighbors.

[0137] Step 7: Sort all data points according to the LOF score, then set a threshold based on business requirements or data characteristics, and identify data points with LOF scores higher than the threshold as outliers.

[0138] The scheme described herein is mainly implemented by adjusting parameters of different algorithms and integrating algorithms of the same type in the model adjustment step.

[0139] (1) Parameter adjustment: The use of algorithms to identify violations often lacks historical data with labels. Therefore, in order to further improve model performance and achieve more accurate identification, reasonable means are needed to adjust model parameters. For the three anomaly detection algorithms, Isolation Forest, One-Class Support Vector Machine, and Local Anomaly Factor, they have strong white-box properties, and the intermediate calculation process and results can be easily visualized. Their distinct statistical characteristics, decision boundaries, decision paths, etc. can be observed. Cross-checking can be achieved through manual observation and statistical tests, thereby correcting the results and counter-complementing the model to adjust parameters and improve performance. For ARIMA, Prophet, and LSTM models, as time series analysis algorithms, model performance can be evaluated through various point-by-point error indicators, and the training-evaluation process can be embedded in the TPE optimization algorithm to achieve automatic optimization.

[0140] (2) Algorithm integration: The results of training and prediction using a single algorithm are often incomplete or inaccurate. Therefore, in order to achieve more accurate outlier detection, a reasonable means is needed to integrate the results of the algorithms. The three anomaly detection algorithms, Isolation Forest, Single-Class Support Vector Machine, and Local Anomaly Factor, have completely different principles, but the results are all to mark abnormal data, so they are easy to integrate. The results are fused through voting methods, weighting methods, etc., to express the degree of abnormality of the abnormal data with more accurate scoring values.

[0141] In summary, in view of the current pain points of orthopedic medical consumables supervision, such as single means, low efficiency, low screening accuracy, and difficulty in mining new models, the hierarchical analysis method of machine learning based on big data can process massive data and mine unknown new patterns of violations by integrating machine learning methods at different levels and with different goals. It can improve the effectiveness and accuracy of the supervision of the use of consumables in medical institutions, and the specific effects are as follows:

[0142] (1) Improve fund utilization efficiency and reduce regulatory costs. The introduction of machine learning hierarchical analysis methods based on big data will significantly improve the ability to identify violations, promote the intelligence level of medical regulatory agencies, reduce fund losses caused by violations, and thus improve fund utilization efficiency. At the same time, with the help of big data and artificial intelligence technology, regulators do not need to have an in-depth understanding of the feature construction process of the algorithm model. They only need to select inspection targets or conduct random exploration to perform modeling analysis and pattern exploration on the current data. The fully automated process has greatly improved the operation and regulatory efficiency of medical funds, effectively reduced the manpower and learning costs of regulatory departments, and improved work efficiency.

[0143] (2) Strengthen supervision and ensure that fund risks are controllable. After the hierarchical analysis method of machine learning based on big data is put into operation, it will enhance the ability to identify and supervise medical fraud. Machine learning algorithm models can better characterize medical transaction behaviors and patterns, making the characteristics of fraud more obvious and easier to be captured by efficient classification models; unsupervised learning can discover more unknown behavior patterns in large amounts of data, transforming from passive supervision to active discovery, and continuously providing new discovery methods for regulatory agencies. The combination of the two can gradually discover existing unknown fraud patterns and adapt to new fraud patterns, effectively combat hidden and diverse fraud methods, actively strengthen the supervision of the medical intelligent audit system, curb the unreasonable growth of medical expenses, and reduce fund operation risks.

[0144] (3) Reduce unnecessary consumption of medical resources and effectively control the irrational growth of medical expenses. By using intelligent technology to identify fraudulent medical behaviors, unnecessary medical expenses can be significantly reduced, thereby limiting the expenditure of medical funds at the overall level;

[0145] (4) Ensure the healthy and sustainable operation of the medical insurance fund. Fraudulent behavior has a serious impact on the balance of income and expenditure of the fund, and intelligent identification technology can ensure the reasonable and standardized use of the fund and safeguard the long-term stable operation of the fund;

[0146] (5) Defend the public’s medical rights and enhance their sense of medical security;

[0147] (6) Enhance the credibility of the medical insurance system. By effectively combating fraud, the public’s trust in the medical system can be enhanced, thereby improving the overall satisfaction of society;

[0148] (7) Improve the mechanisms and procedures of medical supervision. Use intelligent algorithms to optimize the supervision process, reduce supervision costs, and improve supervision efficiency;

[0149] (8) Promote the construction of an industry integrity system and lay a solid foundation for the long-term and healthy development of the industry;

[0150] (9) Have a positive impact on the improvement of laws and regulations. The experience accumulated in intelligent identification of insurance fraud can provide strong support for the revision and improvement of laws and regulations related to medical insurance;

[0151] (10) Strictly prevent and crack down on medical fraud, and form a strong deterrent to potential illegal and irregular behaviors. The anti-fraud model based on deep learning can move the regulatory threshold forward, conduct real-time supervision during the process, effectively identify and crack down on new types of fraud, and have a strong deterrent effect on those who intend to violate the law;

[0152] (11) It has enhanced the fund's sustained protection capabilities, strengthened the public's trust in the medical security system, improved the quality and fairness of medical services, enhanced trust in the medical security system, and promoted social stability and vitality.

[0153] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for identifying illegal use of medical consumables based on big data, characterized in that: The following steps are involved: Step 1: Count the usage of consumables: Count the related indicators of consumables usage and make Pareto analysis to determine the key monitoring list; Step 2, data collection: extracting medical institution, department, physician, consumables, inventory, DRG / DIP information from data sources such as diagnosis and treatment database, regulatory department database, medical institution procurement database, and patient medical record database; Step 3, data cleaning: processing missing values ​​and outliers, completing data format conversion, deduplication, type verification, and standardization; Step 4, feature engineering: select and transform massive data so that its numerical regularity can be easily captured by the algorithm and reduce the impact of noise data on analysis accuracy; Step 5, macro analysis: use a variety of time series analysis methods to do long-term trend analysis; Step 6, meso-level analysis: Use a variety of extreme analysis, cluster analysis and other methods to find out the directionality and focus on the key points; Step 7, micro-analysis: conduct abnormal detection and analysis on each type of consumables on the key monitoring list to find out the suspicious points; Step 8, model adjustment: adjust steps 1 to 7 according to the actual model operation results to improve the model performance in an orderly manner.

2. The method for identifying illegal use of medical consumables based on big data according to claim 1, characterized in that: In the step of counting the use of consumables, the settlement amount of all medical consumables in the past three years is counted, and the Pareto analysis method is used to find out the types of consumables that meet the characteristics of large amounts involved, large usage, and high unit prices. The statistical caliber of the types refers to the three-level classification catalog of medical consumables in the country; After analysis, it was found that in the past three years, the settlement amount of the top 30 medical consumables in the city accounted for 38% of the total settlement amount of all consumables, and formed a Pareto inflection point, thus being included in the key monitoring list.

3. The method for identifying illegal use of medical consumables based on big data according to claim 2, characterized in that: In the data acquisition step, the following data is obtained: (1) The settlement list and detailed items in the medical transaction database are used as the main data; (2) Medical institution information, physician information, natural person information, and disease group information in the medical transaction database are used as dimensional data; (3) Historical law enforcement and penalty information in the regulatory database is used as one of the label data; (4) The consumables procurement and inventory information in the medical institution procurement database serves as the main data.

4. The method for identifying illegal use of medical consumables based on big data according to claim 3 is characterized in that: In the data cleaning step, noise is removed by: (1) Remove missing values: For missing values, review the dictionary value of the field and process it by filling, deleting or interpolating; (2) Duplicate value processing: For the duplicate values ​​of each field, review their meaning and delete necessary records; (3) Data standardization / normalization: For continuous fields, standardization / normalization is performed to make them more consistent with model input requirements; (4) Type conversion: For each field, convert it to a numeric or enumeration type to meet the model input requirements; (5) Character processing: For each character field, delete special symbols, spaces, line breaks, and invalid characters that affect data quality; (6) Consistency check: Perform consistency checks on fields in each form that are used to connect or express the same meaning to ensure that their field names and field dictionary values ​​are consistent.

5. The method for identifying illegal use of medical consumables based on big data according to claim 4 is characterized in that: In the feature engineering step, data features are extracted or enhanced mainly through the following methods: (1) Data discretization: For some continuous fields, which are mainly expressed in segmented values ​​in business, discretization processing is performed; (2) Sample balance: For data with labels, examine the ratio of each type of label and use the SMOTE algorithm to increase or decrease the amount of data corresponding to certain types of labels so that the amount of data corresponding to each type of label is close. (3) Data encoding: Encode all categorical fields, such as natural encoding, One-Hot encoding, etc. (4) Field segmentation: Split the merged fields. For example, split the "year, month, and day" field into "year", "month", and "day".

6. The method for identifying illegal use of medical consumables based on big data according to claim 5, characterized in that: In the macro analysis step, the time series analysis method is used to analyze the usage indicators of various consumables, mainly the usage quantity and cost. The highest statistical caliber of time is month and the lowest is hour. In this way, a time series of indicators ranging from dozens to thousands is formed for each type of consumables. According to the characteristics of each time series, select appropriate analysis methods, including but not limited to the following methods: (1) Autoregressive Integrated Moving Average (ARIMA) model: For time series analysis of indicators with "month" as the statistical caliber, the cycle, fluctuation, steady state and other information are extracted to determine the three parameters p, d and q. After fitting the data, the values ​​of multiple points in the future can be predicted, and the confidence interval is set to no less than 90%; (2) Prophet model: For time series analysis of indicators with "day" as the statistical caliber, the trend, discontinuity points, and fluctuation cycle information are extracted to determine important parameters of the model such as growth, checkpoints, and seasonal_mode. After fitting the data, the values ​​of multiple points in the future can be predicted, and the confidence interval is set to be no less than 90%; (3) LSTM model based on attention mechanism: For the time series analysis of indicators with "hours" as the statistical caliber, the nonlinear correlation between the points in the specified time window is extracted and expressed in the form of normalized weights; Through the multi-head attention mechanism, the changing patterns of consumables indicators at different scales can be accurately captured; Through training, the specific structure of the model, the parameters of each layer, and the parameters of each neuron are determined, and the values ​​of multiple points in the future are predicted, with the confidence interval limited to no less than 90%.

7. The method for identifying illegal use of medical consumables based on big data according to claim 6, characterized in that: In the observation analysis step, the analysis is performed through extreme analysis and clustering algorithms, including: (1) Extreme analysis: Use big data technology to conduct statistical analysis on the number of consumables used, screen cases where the number of one-time uses exceeds the threshold X, and focus on them. If the number of such cases is large, then conduct the next step of cluster analysis. If the number is small, then turn to the correlation analysis of other items prescribed at the same time; (2) Cluster analysis: K-means, DBSCAN, and BIRCH clustering algorithms were used to analyze the directionality, focusing on the three dimensions of medical institutions, doctors, and consumables manufacturers. The analysis was mainly based on statistical indicators such as type, quantity, number of people, and number of visits; When payment is made according to DRG / DIP, additional directional analysis of disease groups is performed to obtain directional analysis in four dimensions: disease groups, medical institutions, doctors, and consumables manufacturers, and key marking is performed.

8. The method for identifying illegal use of medical consumables based on big data according to claim 7, characterized in that: In the micro-analysis step, the unreasonable usage range of various consumables and costs is determined through an outlier analysis method centered on the type of consumables; the algorithms used are mainly isolation forest, single-class support vector machine, and local anomaly factor method; (1) Isolation Forest: An efficient unsupervised learning algorithm for outlier detection. The basic principle is to recursively and randomly select data features and split points to split the data set into multiple subsets until only one data point remains in each subset. Since the number of abnormal data is small and the difference from normal data is large, fewer steps are required when they are isolated, so they can be identified faster. The steps are as follows: Step 1, random initialization: randomly select the number of trees in the isolation forest; Step 2, construct a single isolated number: for each isolated number, randomly select a data point from the data set as the root node of the tree; randomly select a feature and a split value on this feature; divide the data point into two subsets according to the split value, and each subset will serve as the root node of the left and right subtrees; Step 3, for each non-leaf node: repeat Step 2, randomly select features and split values ​​until the stopping condition is met, which is either the preset maximum tree depth or the number of samples in the node is less than a certain threshold; Step 4, repeat Step 1 to Step 3 to complete the construction of multiple trees; Step 5, for each sample point, calculate its depth on each tree, and then calculate the average depth as the outlier score of the sample; outliers usually have shorter paths and therefore smaller depths; Step 6, training: Use training data for training, and judge whether the model training is reasonable and appropriate by statistical anomaly scores, visualization, the structure of isolated trees, and the decision path of isolated trees; Step 7, prediction: assign a prediction label to the new data point, -1 indicates abnormality and 1 indicates normality.

9. (2) One-class support vector machine: an unsupervised learning algorithm widely used in the field of anomaly detection; The core idea is to maximize the distance from normal data points to the decision boundary, similar to maximizing the interval in the support vector machine (SVM), so as to find the normal patterns in the data and mark the data points that are significantly different from these patterns as abnormal. The steps are as follows: Step 1, set the kernel function to be used, usually selected from linear, polynomial, sigmoid and RBF kernels; Step 2, set the parameters n and g, where n is a value between 0 and 1, indicating the allowable misclassification ratio, that is, the expected ratio of normal samples that are misclassified as abnormal, and g defines the influence range of a single training sample, usually used with the Gaussian kernel function; Step 3, training: use the prepared data and set parameters to train the single-class support vector machine model; Step 4, Evaluation: The model calculates a score for each sample, which indicates the distance from the sample to the decision boundary. Usually, samples with scores below a certain threshold are considered abnormal. The decision boundary is observed by visualization to adjust the model parameters to appropriate values. Step 5, determine a threshold to distinguish normal and abnormal samples: this threshold can make a certain proportion of known normal samples be classified as normal scores as the threshold; (3) Local anomaly factor: A density-based unsupervised anomaly detection method. The core idea is to identify anomalies by comparing the local density deviation between a data point and its neighbors. If the local density of a data point is significantly lower than that of its neighbors, then the point may be an anomaly. The density ratio of the data point to its k nearest neighbors is calculated to determine the anomaly score. This ratio is called the local anomaly factor. The steps are as follows: Step 1, select the number of neighbors parameter k: Determine a parameter k, which represents the k nearest neighbors of each data point; this parameter is one of the most important hyperparameters in the algorithm and needs to be adjusted according to the specific data set; Step 2, calculate K distance: For each data point P, calculate its distance to the kth neighbor among its k nearest neighbors. This distance is called the K distance of point P. Step 3, determine the K neighborhood: use the K distance to determine the K neighborhood of each data point, that is, the set of all points whose distance from the data point P is not greater than K distance; Step 4, calculate the reachable distance RD: For each neighbor O of the data point P, calculate the reachable distance from O to P, defined as max(K-distance(O), dist(P, O)), where dist(P, O) is the actual distance between P and O; Step 5, calculate the local reachability density LRD: For each data point P, calculate the inverse of the average of the reachability distances of all its neighbors as the local reachability density of the point; the local reachability density reflects the local density of point P; Step 6, calculate the local anomaly factor LOF: For each data point P, calculate its LOF score, which is the average of the ratios of the local reachability density of its k nearest neighbors to the local reachability density of P. This score reflects the degree of anomaly of point P relative to its neighbors; Step 7: Sort all data points according to the LOF score, then set a threshold based on business requirements or data characteristics, and identify data points with LOF scores higher than the threshold as outliers.

10. The method for identifying illegal use of medical consumables based on big data according to claim 8, characterized in that: In the model adjustment step, the following is achieved by adjusting the parameters of different algorithms and integrating algorithms of the same type: (1) Parameter adjustment: When using algorithms to identify violations, there is often a lack of historical data with labels. For the three anomaly detection algorithms, Isolation Forest, Single-Class Support Vector Machine, and Local Anomaly Factor, they have strong white-box properties, and the intermediate calculation process and results are easy to visualize. Their distinct statistical characteristics, decision boundaries, and decision paths can be observed. Cross-checking can be achieved through manual observation and statistical testing, thereby correcting the results, counter-complementing the model to adjust parameters, and improving performance. For ARIMA, Prophet, and LSTM models, as time series analysis algorithms, various point-by-point error indicators are used to evaluate model performance, and the training-evaluation process is embedded in the TPE optimization algorithm to achieve automatic optimization; (2) Algorithm integration: The results of training and prediction using a single algorithm are often incomplete or inaccurate. The three anomaly detection algorithms, isolation forest, single-class support vector machine, and local anomaly factor, have completely different principles, but the results are all to mark abnormal data, so they are integrated; the results are fused through voting and weighting methods.